Re: [PATCH RFC v2 1/4] cgroup: implement eventfd-based generic API for notifications

[Kirill A. Shutemov]

On Tue, Dec 15, 2009 at 11:35 AM, KAMEZAWA Hiroyuki
<kamezawa.hiroyu@xxxxxxxxxxxxxx> wrote:
> On Tue, 15 Dec 2009 11:11:16 +0200
> "Kirill A. Shutemov" <kirill@xxxxxxxxxxxxx> wrote:
>
>> Could anybody review the patch?
>>
>> Thank you.
>
> some nitpicks.
>
>>
>> On Sat, Dec 12, 2009 at 12:59 AM, Kirill A. Shutemov
>> <kirill@xxxxxxxxxxxxx> wrote:
>
>> > + Â Â Â /*
>> > + Â Â Â Â* Unregister events and notify userspace.
>> > + Â Â Â Â* FIXME: How to avoid race with cgroup_event_remove_work()
>> > + Â Â Â Â* Â Â Â Âwhich runs from workqueue?
>> > + Â Â Â Â*/
>> > + Â Â Â mutex_lock(&cgrp->event_list_mutex);
>> > + Â Â Â list_for_each_entry_safe(event, tmp, &cgrp->event_list, list) {
>> > + Â Â Â Â Â Â Â cgroup_event_remove(event);
>> > + Â Â Â Â Â Â Â eventfd_signal(event->eventfd, 1);
>> > + Â Â Â }
>> > + Â Â Â mutex_unlock(&cgrp->event_list_mutex);
>> > +
>> > +out:
>> > Â Â Â Âreturn ret;
>> > Â}
>
> How ciritical is this FIXME ?

There is potential race. I have never seen it. When userspace closes
eventfd associated
with cgroup event, cgroup_event_remove() will not be called
immediately. It will be called
later from workqueue. If somebody removes cgroup before the workqueue calls
cgroup_event_remove() we will get problem.
It's unlikely, but theoretically possible.

> But Hmm..can't we use RCU ?

I'll play with it.

>> >
>> > @@ -1136,6 +1187,8 @@ static void init_cgroup_housekeeping(struct cgroup *cgrp)
>> > Â Â Â ÂINIT_LIST_HEAD(&cgrp->release_list);
>> > Â Â Â ÂINIT_LIST_HEAD(&cgrp->pidlists);
>> > Â Â Â Âmutex_init(&cgrp->pidlist_mutex);
>> > + Â Â Â INIT_LIST_HEAD(&cgrp->event_list);
>> > + Â Â Â mutex_init(&cgrp->event_list_mutex);
>> > Â}
>> >
>> > Âstatic void init_cgroup_root(struct cgroupfs_root *root)
>> > @@ -1935,6 +1988,16 @@ static const struct inode_operations cgroup_dir_inode_operations = {
>> > Â Â Â Â.rename = cgroup_rename,
>> > Â};
>> >
>> > +/*
>> > + * Check if a file is a control file
>> > + */
>> > +static inline struct cftype *__file_cft(struct file *file)
>> > +{
>> > + Â Â Â if (file->f_dentry->d_inode->i_fop != &cgroup_file_operations)
>> > + Â Â Â Â Â Â Â return ERR_PTR(-EINVAL);
>> > + Â Â Â return __d_cft(file->f_dentry);
>> > +}
>> > +
>> > Âstatic int cgroup_create_file(struct dentry *dentry, mode_t mode,
>> > Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Âstruct super_block *sb)
>> > Â{
>> > @@ -2789,6 +2852,151 @@ static int cgroup_write_notify_on_release(struct cgroup *cgrp,
>> > Â Â Â Âreturn 0;
>> > Â}
>> >
>> > +static inline void cgroup_event_remove(struct cgroup_event *event)
>> > +{
>> > + Â Â Â struct cgroup *cgrp = event->cgrp;
>> > +
>> > + Â Â Â BUG_ON(event->cft->unregister_event(cgrp, event->cft, event->eventfd));
>
> Hmm ? BUG ? If bug, please add document or comment.

I'll remove it, since we check it in cgroup_write_event_control().

>> > + Â Â Â eventfd_ctx_put(event->eventfd);
>> > + Â Â Â remove_wait_queue(event->wqh, &event->wait);
>> > + Â Â Â list_del(&event->list);
>
> please add comment as /* event_list_mutex must be held */

Ok.

>> > + Â Â Â kfree(event);
>> > +}
>> > +
>> > +static void cgroup_event_remove_work(struct work_struct *work)
>> > +{
>> > + Â Â Â struct cgroup_event *event = container_of(work, struct cgroup_event,
>> > + Â Â Â Â Â Â Â Â Â Â Â remove);
>> > + Â Â Â struct cgroup *cgrp = event->cgrp;
>> > +
>> > + Â Â Â mutex_lock(&cgrp->event_list_mutex);
>> > + Â Â Â cgroup_event_remove(event);
>> > + Â Â Â mutex_unlock(&cgrp->event_list_mutex);
>> > +}
>> > +
>> > +static int cgroup_event_wake(wait_queue_t *wait, unsigned mode,
>> > + Â Â Â Â Â Â Â int sync, void *key)
>> > +{
>> > + Â Â Â struct cgroup_event *event = container_of(wait,
>> > + Â Â Â Â Â Â Â Â Â Â Â struct cgroup_event, wait);
>> > + Â Â Â unsigned long flags = (unsigned long)key;
>> > +
>> > + Â Â Â if (flags & POLLHUP)
>> > + Â Â Â Â Â Â Â /*
>> > + Â Â Â Â Â Â Â Â* This function called with spinlock taken, but
>> > + Â Â Â Â Â Â Â Â* cgroup_event_remove() may sleep, so we have
>> > + Â Â Â Â Â Â Â Â* to run it in a workqueue.
>> > + Â Â Â Â Â Â Â Â*/
>> > + Â Â Â Â Â Â Â schedule_work(&event->remove);
>> > +
>> > + Â Â Â return 0;
>> > +}
>
>> > +
>> > +static void cgroup_event_ptable_queue_proc(struct file *file,
>> > + Â Â Â Â Â Â Â wait_queue_head_t *wqh, poll_table *pt)
>> > +{
>> > + Â Â Â struct cgroup_event *event = container_of(pt,
>> > + Â Â Â Â Â Â Â Â Â Â Â struct cgroup_event, pt);
>> > +
>> > + Â Â Â event->wqh = wqh;
>> > + Â Â Â add_wait_queue(wqh, &event->wait);
>> > +}
>> > +
>> > +static int cgroup_write_event_control(struct cgroup *cont, struct cftype *cft,
>> > + Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â const char *buffer)
>> > +{
>> > + Â Â Â struct cgroup_event *event = NULL;
>> > + Â Â Â unsigned int efd, cfd;
>> > + Â Â Â struct file *efile = NULL;
>> > + Â Â Â struct file *cfile = NULL;
>> > + Â Â Â char *endp;
>> > + Â Â Â int ret;
>> > +
>> > + Â Â Â efd = simple_strtoul(buffer, &endp, 10);
>> > + Â Â Â if (*endp != ' ')
>> > + Â Â Â Â Â Â Â return -EINVAL;
>> > + Â Â Â buffer = endp + 1;
>> > +
>> > + Â Â Â cfd = simple_strtoul(buffer, &endp, 10);
>> > + Â Â Â if ((*endp != ' ') && (*endp != '\0'))
>> > + Â Â Â Â Â Â Â return -EINVAL;
>> > + Â Â Â buffer = endp + 1;
>> > +
>> > + Â Â Â event = kzalloc(sizeof(*event), GFP_KERNEL);
>> > + Â Â Â if (!event)
>> > + Â Â Â Â Â Â Â return -ENOMEM;
>> > + Â Â Â event->cgrp = cont;
>> > + Â Â Â INIT_LIST_HEAD(&event->list);
>> > + Â Â Â init_poll_funcptr(&event->pt, cgroup_event_ptable_queue_proc);
>> > + Â Â Â init_waitqueue_func_entry(&event->wait, cgroup_event_wake);
>> > + Â Â Â INIT_WORK(&event->remove, cgroup_event_remove_work);
>> > +
>> > + Â Â Â efile = eventfd_fget(efd);
>> > + Â Â Â if (IS_ERR(efile)) {
>> > + Â Â Â Â Â Â Â ret = PTR_ERR(efile);
>> > + Â Â Â Â Â Â Â goto fail;
>> > + Â Â Â }
>> > +
>> > + Â Â Â event->eventfd = eventfd_ctx_fileget(efile);
>> > + Â Â Â if (IS_ERR(event->eventfd)) {
>> > + Â Â Â Â Â Â Â ret = PTR_ERR(event->eventfd);
>> > + Â Â Â Â Â Â Â goto fail;
>> > + Â Â Â }
>> > +
>> > + Â Â Â cfile = fget(cfd);
>> > + Â Â Â if (!cfile) {
>> > + Â Â Â Â Â Â Â ret = -EBADF;
>> > + Â Â Â Â Â Â Â goto fail;
>> > + Â Â Â }
>> > +
>> > + Â Â Â /* the process need read permission on control file */
>> > + Â Â Â ret = file_permission(cfile, MAY_READ);
>> > + Â Â Â if (ret < 0)
>> > + Â Â Â Â Â Â Â goto fail;
>> > +
>> > + Â Â Â event->cft = __file_cft(cfile);
>> > + Â Â Â if (IS_ERR(event->cft)) {
>> > + Â Â Â Â Â Â Â ret = PTR_ERR(event->cft);
>> > + Â Â Â Â Â Â Â goto fail;
>> > + Â Â Â }
>> > +
>> > + Â Â Â if (!event->cft->register_event || !event->cft->unregister_event) {
>> > + Â Â Â Â Â Â Â ret = -EINVAL;
>> > + Â Â Â Â Â Â Â goto fail;
>> > + Â Â Â }
>> > +
>> > + Â Â Â ret = event->cft->register_event(cont, event->cft,
>> > + Â Â Â Â Â Â Â Â Â Â Â event->eventfd, buffer);
>> > + Â Â Â if (ret)
>> > + Â Â Â Â Â Â Â goto fail;
>> > +
>> > + Â Â Â efile->f_op->poll(efile, &event->pt);
>
> Not necessary to check return value ?

You are right. We need to check return value for POLLHUP.

Thanks!
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/